<p>This paper develops a theoretical and empirical framework for understanding AI-generated images through Gilbert Simondon's philosophy of individuation. Against approaches that treat generative AI outputs as either unprecedented creations or recombinations of training data, we argue that AI image generation constitutes a process of ontogenesis: the coming-into-being of individuals from a preindividual field of potentials. Drawing on Simondon's concepts of metastability, transduction, and individuation, we analyze the latent space of diffusion models as a structured field of tensions whose resolution produces individuated images. Our methodology combines computational experiments on Stable Diffusion 2.1 with three artistic case studies (Refik Anadol, Sougwen Chung, Holly Herndon) to demonstrate how Simondonian concepts illuminate both the technical operations and aesthetic dimensions of AI image generation. Empirically, we show that latent space exhibits metastable structure (effective dimensionality compressed to a participation ratio of 52.9, roughly ninefold below the isotropic prior), that denoising trajectories follow transductive dynamics with identifiable phase transitions, and that variations from identical prompts display structured semantic coherence (mean intra-prompt CLIP similarity 0.827, intra-prompt vs. intra-category d = 1.385) inconsistent with the 'error' interpretation. We propose that what are commonly termed 'hallucinations' are better understood as alternative individuations: legitimate resolutions of preindividual potentials rather than failures. This reconceptualization has implications for debates about AI creativity, authorship, and the ontological status of machine-generated artifacts. By mapping the ontogenesis of AI images, we contribute to emerging dialogues between philosophy of technology and computational practice, offering a 'technical philosophy' that neither reduces AI to mechanism nor inflates it into artificial consciousness.</p>